Treatment-Confounder Feedback

Why Confounding Can Change Over Time

In a longitudinal study, treatment decisions are often repeated. A clinician may start a treatment, observe a patient's changing condition, and then decide whether to continue, increase, or stop treatment. The changing condition can be important for the outcome and can guide the next treatment decision. At the same time, it may itself have been changed by the earlier treatment.

This creates treatment-confounder feedback: an earlier treatment affects a later covariate, and that covariate helps determine subsequent treatment while also predicting the outcome. The word “feedback” refers to this sequence across time. It does not mean that treatment and covariate are determined simultaneously, or that a later event causes an earlier one.

Definition: Treatment-confounder feedback is a longitudinal pattern in which treatment at an earlier time affects a covariate measured later, and that later covariate is a confounder for a subsequent treatment-outcome relationship. In the two-decision case, \(A_0\) affects \(L_1\), while \(L_1\) influences both the later treatment \(A_1\) and the outcome \(Y\).

Here \(A_0\) and \(A_1\) denote treatment at the first and second decision times, \(L_1\) denotes a covariate measured after \(A_0\) but before \(A_1\), and \(Y\) is measured after the treatment sequence. A covariate is a confounder for a particular treatment decision when it is a cause or relevant predictor of both that treatment and the outcome, given the preceding history. Thus \(L_1\) can be a confounder for \(A_1\) even though it is not a baseline covariate.

The Time Ordering and Potential Outcomes

The order of measurement is essential. First \(A_0\) is assigned; then \(L_1\) is measured; then \(A_1\) is assigned; and later \(Y\) is observed. The later treatment cannot be used to explain the earlier measurement of \(L_1\). By contrast, \(A_0\) can affect \(L_1\), because it comes first.

Using potential-outcome notation for treatment sequences, \(L_1^{a_0}\) denotes the value the intermediate covariate would take under first treatment \(a_0\), and \(Y^{a_0,a_1}\) denotes the outcome under the treatment sequence \((a_0,a_1)\). Feedback is possible when the distribution of \(L_1^{a_0}\) changes with \(a_0\), and the value of \(L_1\) in turn influences which \(A_1\) is received and the later outcome. These statements describe a causal structure; they do not by themselves identify a causal effect from observed data.

A useful way to recognize the pattern is to ask three separate questions at the second decision: Was the covariate measured before \(A_1\)? Can \(A_0\) change it? Does it help determine \(A_1\) and predict \(Y\)? If all three answers are yes, the covariate participates in treatment-confounder feedback.

Worked Example: Disease Severity Before a Second Treatment Decision

Suppose a hypothetical clinic records an initial treatment \(A_0\), then classifies disease severity \(L_1\) as severe (\(1\)) or not severe (\(0\)), and then decides whether to give a second treatment \(A_1\). Assume the following assignment and covariate probabilities:

P(L_1=1\mid A_0=0)=0.20,\qquad P(L_1=1\mid A_0=1)=0.70.

For this example, the later-treatment mechanism is explicitly assumed to depend on \(L_1\) alone: \(P(A_1=1\mid L_1=0,A_0)=0.30\) and \(P(A_1=1\mid L_1=1,A_0)=0.80\). In particular, the conditional probability is the same at both values of \(A_0\) once \(L_1\) is specified.

For patients with \(A_0=0\), the total-probability calculation is:

\begin{aligned} P(A_1=1\mid A_0=0) &=P(A_1=1\mid L_1=1,A_0=0)P(L_1=1\mid A_0=0)\\ &\quad+P(A_1=1\mid L_1=0,A_0=0)P(L_1=0\mid A_0=0)\\ &=(0.80)(0.20)+(0.30)(0.80)\\ &=0.16+0.24=0.40. \end{aligned}

For patients with \(A_0=1\), the same calculation gives:

\begin{aligned} P(A_1=1\mid A_0=1) &=(0.80)(0.70)+(0.30)(0.30)\\ &=0.56+0.09=0.65. \end{aligned}

The probability of the second treatment differs across initial-treatment groups because the groups have different distributions of severity, and severity strongly guides the second decision. If severity also predicts the later outcome, it is a time-varying confounder for the \(A_1\)-outcome relationship. The example illustrates the feedback structure; the probabilities alone do not establish a causal effect of either treatment.

Why Conditioning on the Full History Matters

The general probability identity for the second treatment conditions on both the intermediate covariate and the preceding treatment. For binary \(L_1\), it is:

\begin{aligned} P(A_1=1\mid A_0=a) &=\sum_{l\in\{0,1\}}P(A_1=1\mid L_1=l,A_0=a)\\ &\qquad\qquad\cdot P(L_1=l\mid A_0=a). \end{aligned}

This is the law of total probability applied within the group with \(A_0=a\). The term \(P(A_1=1\mid L_1=l,A_0=a)\) must retain \(A_0=a\) unless the treatment-assignment mechanism justifies dropping it. For example, one may assume that \(A_1\) is assigned based only on \(L_1\), so that \(P(A_1=1\mid L_1,A_0)=P(A_1=1\mid L_1)\). Without that assumption, replacing the full conditional probability with \(P(A_1=1\mid L_1)\) is not valid in general.

This point is substantive, not merely notational. Prior treatment can continue to matter for later treatment even after accounting for the measured covariate—for example, because the clinician considers treatment history. A correct calculation represents the assignment process actually specified, rather than silently imposing a simpler one.

Worked Example: Later Treatment Depends on Both Severity and Prior Treatment

Consider a second hypothetical clinic. Let \(L_1=1\) indicate a severe condition. Suppose the severity probabilities are \(P(L_1=1\mid A_0=0)=0.35\) and \(P(L_1=1\mid A_0=1)=0.65\). In this clinic, the later-treatment probabilities depend on both severity and prior treatment:

\begin{aligned} P(A_1=1\mid L_1=0,A_0=0)&=0.20,& P(A_1=1\mid L_1=1,A_0=0)&=0.70,\\ P(A_1=1\mid L_1=0,A_0=1)&=0.30,& P(A_1=1\mid L_1=1,A_0=1)&=0.80. \end{aligned}

For \(A_0=0\), the full-history calculation is:

\begin{aligned} P(A_1=1\mid A_0=0) &=(0.70)(0.35)+(0.20)(0.65)\\ &=0.245+0.130=0.375. \end{aligned}

For \(A_0=1\), it is:

\begin{aligned} P(A_1=1\mid A_0=1) &=(0.80)(0.65)+(0.30)(0.35)\\ &=0.520+0.105=0.625. \end{aligned}

These calculations use the appropriate conditional treatment probability for each value of \(A_0\). Assuming instead that \(A_1\) depends on severity alone would replace the two treatment probabilities for each severity level with a single value; the example provides no basis for doing so.

A Short Derivation of the Feedback Link

The two-stage mechanism can be summarized without assuming that the later treatment depends on the intermediate covariate alone. Write \(q(l,a)=P(A_1=1\mid L_1=l,A_0=a)\) and \(r(l\mid a)=P(L_1=l\mid A_0=a)\). Then:

P(A_1=1\mid A_0=a)=\sum_l q(l,a)r(l\mid a).

To obtain this expression, partition the event \(A_1=1\) into the mutually exclusive cases \(L_1=l\), within the group \(A_0=a\). For each case, its contribution to the probability is the chance of receiving \(A_1=1\) in that case multiplied by the chance of being in that case. Adding the contributions gives the displayed sum. If the earlier treatment changes \(r(l\mid a)\), then the distribution of the intermediate covariate changes across earlier-treatment groups. If \(q(l,a)\) also varies with \(l\), that changed distribution can alter the probability of later treatment. This is the mathematical route by which an earlier treatment can shape a later treatment decision through \(L_1\).

The derivation does not say that every observed difference in later treatment is caused by \(A_0\). It states how the specified conditional distributions combine. A causal interpretation of a comparison between treatment histories requires appropriate design and identification assumptions, which are separate from recognizing the feedback structure.

Worked Example: A Biomarker That Guides Dose Changes

Suppose a hypothetical study measures a biomarker after an initial dose. Let \(A_0=1\) indicate an initial dose and \(L_1=1\) indicate a high biomarker value. Assume the high-biomarker probabilities are 0.25 for \(A_0=0\) and 0.60 for \(A_0=1\). A clinician gives a subsequent dose \(A_1=1\) with probability 0.75 if the biomarker is high and 0.25 if it is not high. For this example, assume this later-dose mechanism depends on the biomarker alone, conditional on the stated history.

Among those with \(A_0=0\), the chance of a subsequent dose is:

P(A_1=1\mid A_0=0) =(0.75)(0.25)+(0.25)(0.75) =0.1875+0.1875=0.375.

Among those with \(A_0=1\), it is:

P(A_1=1\mid A_0=1) =(0.75)(0.60)+(0.25)(0.40) =0.450+0.100=0.550.

The initial dose is associated with a different biomarker distribution, and the biomarker guides the next dose. If the biomarker also predicts the eventual outcome, it has the defining role of a treatment-affected confounder for the second decision. The calculation is conditional on the stated mechanisms and does not establish that the initial dose caused the difference; in this hypothetical setup, the probabilities are stipulated to explain the structure.

What Feedback Does—and Does Not—Mean

Treatment-confounder feedback is a description of how variables relate over a treatment sequence. It does not require every component to be affected: an earlier treatment may have no effect on some later covariates, and some covariates may predict the outcome without influencing treatment. Nor does it mean the intermediate covariate is necessarily measured without error or that all relevant confounders have been observed.

The key distinction is relative to treatment time. \(L_1\) occurs after \(A_0\), so it can be affected by \(A_0\). It occurs before \(A_1\), so it can help determine \(A_1\). If it also predicts \(Y\), it is relevant to confounding at the later treatment decision. A baseline variable, by contrast, is measured before the first treatment and cannot be a consequence of that treatment.

This structure matters because a single label such as “confounder” can hide when the covariate was measured and which treatment decision it concerns. Recording the temporal order makes clear that the role of a covariate can depend on the decision being studied. Recognizing this pattern is a prerequisite for choosing and interpreting an analysis of treatment sequences; the pattern itself does not dictate a particular estimator.

Common Mistakes

Ignoring the prior treatment in a conditional probability. The general term is \(P(A_1=1\mid L_1,A_0)\). Replacing it with \(P(A_1=1\mid L_1)\) requires a stated assignment assumption or evidence supporting that equality.
Calling every post-treatment variable a confounder. A later covariate is a confounder for a subsequent treatment-outcome relationship only if it is relevant to both the later treatment and outcome, given the history.
Reversing the time order. A covariate measured before \(A_1\) may guide \(A_1\), but \(A_1\) cannot cause that earlier measurement.
Treating feedback as proof of causation from observed associations. The structure identifies a possible causal ordering; observed conditional probabilities alone do not establish the effects of treatment.

Key takeaway: Treatment-confounder feedback occurs when earlier treatment changes a later covariate that influences both subsequent treatment and outcome. Describe the temporal order explicitly, and condition later-treatment probabilities on the full prior history unless a justified assumption allows a simplification.

Check Your Understanding

Answer each question by attending to both the time order and the stated probability model.

In a two-decision study, what ordering must hold for \(L_1\) to be affected by \(A_0\) and still influence \(A_1\)?
Suppose \(L_1\) is affected by \(A_0\) and predicts \(Y\), but does not influence \(A_1\). Does it meet the definition of a confounder for the \(A_1\)-outcome relationship? Explain.
Write the law-of-total-probability expression for \(P(A_1=1\mid A_0=a)\) when \(L_1\) is binary, retaining all relevant conditioning.
What assumption permits replacing \(P(A_1=1\mid L_1,A_0)\) with \(P(A_1=1\mid L_1)\)?
Why does a difference in \(P(A_1=1\mid A_0)\) by itself not prove a causal effect of \(A_0\)?

What You'll Learn